Pareto Feature Selection for the Transfection–Clearance Trade-off in Lipid Nanoparticles

Lipid nanoparticles (LNPs) — tiny fatty particles that carry messenger RNA into cells — face a built-in conflict. To deliver their cargo they must be chemically aggressive: their ionizable lipids must protonate in the acidic environment of a cell's recycling compartment and rupture its membrane. But aggressive, charged, protein-attracting chemistry is exactly what flags a particle for destruction by the reticuloendothelial system (RES), the body's network of scavenger macrophages, dominated by Kupffer cells in the liver, which pull circulating particles out of the blood within minutes. A clinically useful LNP must therefore do two opposing things at once: transfect target cells and avoid being eaten. Current machine learning for LNP design ignores this. The two leading predictors each model a single scalar outcome: LANTERN predicts transfection efficiency from molecular descriptors, and the AGILE platform predicts cell-specific potency — and AGILE's own findings show the endpoints conflict, since the lipid that best delivers to muscle (H9, a 7.8-fold improvement over the benchmark lipid MC3) is not the one that best enters macrophages (R6, a five-fold improvement).

This proposal builds a feature-selection method that treats the conflict itself as the object of study. The idea in two sentences: from a pool of molecular descriptors, use noise-aware Pareto optimization to find the smallest descriptor subsets that jointly predict target-cell transfection and macrophage uptake, and cut the resulting search cost with lazy, cached evaluations so that the optimization never becomes prohibitively expensive. The scientific payoff is an interpretable answer to which physicochemical properties drive delivery versus clearance — the tension that Engineering Ionizable Lipid Nanoparticles for mRNA Delivery describes mechanistically but that no computational model currently maps.

Core Claim

The claim is that Pareto-based multi-objective feature selection, made noise-aware and made cheap, will recover small descriptor subsets that predict both LNP endpoints at accuracy comparable to or better than the strongest published baseline, while exposing which descriptors are shared between the two endpoints and which are genuinely in conflict. Two existing method families supply the machinery and have never met this domain. POSS formulates subset selection as simultaneously minimizing prediction error and subset size, with a proven approximation guarantee of $1 - e^{-\gamma}$ (where $\gamma$ is the submodularity ratio of the objective), and empirically beats greedy forward regression, Lasso-based selection, and recursive feature elimination across 12 regression benchmarks, finding better solutions in only 14–23% of its theoretical runtime bound (Subset Selection by Pareto Optimization). Its successor PONSS adds $\theta$-domination and buffered re-evaluation precisely because noisy evaluations destroy greedy selection, whose approximation ratio collapses to $\Theta(1/k)$ under multiplicative noise (Subset Selection under Noise) — and LNP screens are exactly such noisy evaluations, since AGILE's 1,200-lipid training library was synthesized unpurified, with side-products contaminating labels. But POSS and PONSS have only ever been run on UCI/LIBSVM datasets, never on molecular descriptors. Conversely, LANTERN shows that in this ~1,000-sample regime, a simple multi-layer perceptron on engineered features (Morgan fingerprints plus expert physicochemical descriptors) reaches an $R^2$ of 0.8161 while the deep graph network behind AGILE reaches only 0.2655 on the same curated test set — yet LANTERN treats feature sets as fixed, statically concatenated blocks and models one endpoint. Neither line of work does what this proposal does.

Formally, we select a subset $S$ of descriptors from a pool of $n$ candidates by solving

$$\min_{S \subseteq \{1,\dots,n\}} \; \big( f_{\text{target}}(S),\; f_{\text{MPS}}(S),\; |S| \big),$$

where $f_{\text{target}}(S)$ is the prediction error of an MLP restricted to descriptors in $S$ on target-cell transfection, $f_{\text{MPS}}(S)$ is the analogous error on macrophage uptake (a laboratory proxy for RES clearance), and $|S|$ is the subset size. The output is not a single subset but a front of non-dominated ones, from small-and-crude to large-and-accurate, plus — by running the same procedure per endpoint — two endpoint-specific subsets whose overlap and divergence constitute the result.

Research Program

Three things get built. First, a curated dual-endpoint dataset: the 1,100-lipid benchmark that LANTARN cleaned (correcting 235 label mismatches and 100 duplicates arising from neglected cis/trans tail isomerism) supplies transfection labels, and the AGILE library's paired measurements in HeLa cells and RAW 264.7 macrophages supply the transfection-versus-macrophage-uptake contrast. An honest dependency: if AGILE's paired macrophage labels cannot be obtained, the macrophage endpoint degrades and the project reduces to single-endpoint Pareto selection, which is materially less interesting; the first milestone is therefore securing or reconstructing that paired data, and the fallback is to train the macrophage head on the subset of AGILE lipids with reported RAW 264.7 values and quantify the label scarcity. Second, the optimizer itself, which we call Lazy-PONSS: PONSS's $\theta$-domination and buffered re-evaluation applied to the three-objective problem above, where each candidate evaluation means training a small MLP under cross-validation. Third, the lazy layer that answers the overhead problem. Vanilla PONSS on this problem needs on the order of $O(k^2 n)$ subset evaluations (POSS's worst case), each multiplied by the buffered re-evaluations that noise handling demands, and each evaluation is a model training — the same compounding cost that A Systematic Survey of Automatic Prompt Optimization Techniques formalizes as $O(M \times N \times T)$ in iterative black-box search. Our remedy follows the pattern proven by Lazy Greedy Hypervolume Subset Selection, which caches an optimistic bound on each candidate's contribution in a priority queue and re-evaluates only the current leader, cutting runtime by 91–99% on DTLZ benchmarks without losing solution quality — except it has never been applied outside synthetic test suites. We cache, for every candidate subset, an optimistic quality bound derived from a single once-trained full-feature model (a cheap stand-in ranking, in the spirit of InstructZero's use of a small proxy model in place of an expensive black box), and train the real MLP only when the bound leaves the candidate competitive with the current front. We aim for an order-of-magnitude reduction in trainings at equal front quality; we do not promise LGI-HSS's 99% on real data, since molecular error surfaces lack the strict submodularity that lazy greedy exploits — POSS's guarantee itself degrades gracefully through the ratio $\gamma$ rather than failing outright, and we will report the empirically measured submodularity of our objectives rather than assume it.

The interpretive analysis then compares the target-only front, the macrophage-only front, and the joint front. The hypothesis, grounded in the clearance literature — where membrane fluidity governs Kupffer uptake in Mononuclear phagocyte system blockade and albumin recruitment blocks ApoE-mediated liver uptake in Albumin-recruiting lipid nanoparticle — is that headgroup ionizability descriptors (pKa proxies) are shared across both endpoints while tail-packing and lipophilicity descriptors diverge. That comparison is the deliverable a formulation chemist can act on, and it feeds directly into candidate-level work: lipids on the transfection-versus-clearance Pareto front become the shortlist for synthesis, for which the Synthesis-DAG generative route of A Deep Generative Model for the Design of Synthesizable Ionizable Lipids — which achieves a 92.6% valid lipid rate — guarantees accessibility, replacing that paper's single-objective AGILE-score reward with our joint front.

Baselines

The do-nothing reference is the full-feature MLP on the complete concatenated descriptor block, exactly as LANTERN reports it: an MLP on Morgan fingerprints plus expert physicochemical descriptors, which achieves $R^2 = 0.8161$ on the curated 1,100-lipid benchmark (LANTERN). Any feature-selection method must justify removing features by matching or approaching this number with far fewer inputs; a random subset of the same cardinality as ours gives the floor. The strongest existing competitor is that same LANTERN configuration, because it is the best published predictor on the exact benchmark we use, and it already embodies the field's implicit answer to feature choice (use everything, in two static blocks). We run it under identical data splits, the same MLP architecture and tuning budget, and the same compute; beating it means reaching within its confidence interval on $R^2$ — or exceeding it — with a substantially smaller subset while also matching it on the macrophage endpoint, which it does not attempt at all. The strongest algorithmic competitor within feature selection is greedy forward regression with orthogonal matching pursuit, since it is the standard cheap alternative; POSS already beat both on 12 datasets at $k = 8$ (Subset Selection by Pareto Optimization), but never on molecular data, so we rerun them head-to-head here under the same wrapper and budget. Finally, the key ablation isolates the two added ingredients separately: Lazy-PONSS versus plain PONSS with identical random seeds and acceptance rules (measuring trainings saved at equal front quality, the lazy layer's only job), and PONSS versus POSS versus greedy under controlled label noise (measuring front degradation, the noise layer's only job).

Why not just…?

Why not just use greedy forward selection? It is simpler and cheaper, and if the descriptor pool is nearly independent it would work. The evidence against it here is direct: greedy is blind to synergistic pairs — a tail unsaturation and a headgroup pKa proxy that each predict nothing but jointly predict a lot — and under multiplicative evaluation noise its approximation ratio provably collapses to $\Theta(1/k)$ (Subset Selection under Noise). AGILE's unpurified library is a documented source of exactly that kind of label noise, so the failure mode is not hypothetical in this domain. Our noise-injection experiment will show the degradation curve directly, and if greedy matches PONSS on the clean split, we will report that honestly.

Why not just use Lasso or elastic-net regularization? It folds the accuracy-versus-sparsity trade-off into one differentiable loss, at a fraction of the complexity. But scalarization introduces a fragile penalty weight and, more importantly, cannot reach non-convex regions of the trade-off front; POSS achieved statistically significantly higher $R^2$ than Lasso-based and recursive-elimination baselines on the classical benchmarks (Subset Selection by Pareto Optimization). No reported result found on Lacuna for Lasso's accuracy on the LANTERN lipid benchmark, so we will run it ourselves as a scalarized three-objective variant (two endpoint errors plus an $L_1$ penalty) with the same tuning budget. If Lasso matches our front on this data, the honest conclusion is that the Pareto machinery matters only under noise and synergy, and the experiments are designed to reveal which.

Why not just train a big graph neural network end-to-end and skip feature selection? This is the most natural ML objection, and it is settled empirically rather than by argument: on the curated lipid benchmark, learned graph embeddings underperformed dramatically, with the AGILE deep framework reaching $R^2 = 0.2655$ versus 0.8161 for the simple MLP on engineered features, a gap LANTERN attributes to over-smoothing and to the ~1,000-sample regime being too small to resolve even cis/trans isomerism. End-to-end learning also cannot deliver the interpretability that is the point of the exercise: a GNN that predicts both endpoints tells you nothing about which properties conflict.

Why not just co-administer decoy particles to saturate the macrophages? This is a real alternative because it attacks clearance without touching formulation: pre-injecting fluid DOPC:cholesterol liposomes extends co-delivered carriers' half-life 2.1-fold, from 1.1 to 2.3 hours (Mononuclear phagocyte system blockade), and sequential "Engage & Evasion" dosing achieves similar escape for extracellular vesicles (Engage & Evasion). It is a genuine threat and we do not argue it away: decoys optimize the administration protocol while we optimize the vehicle, and the two are complementary. But decoy saturation transiently disables an innate immune filter with unknown infection and toxicity risk, requires rigid multi-hour dosing intervals, and — decisively for our claims — does nothing for target-cell transfection, since a rescued particle still needs its own protonation-driven membrane rupture to deliver cargo (Engineering Ionizable Lipid Nanoparticles for mRNA Delivery). Our feature analysis is precisely the tool that would tell you when a decoy is unnecessary because the formulation itself is clearance-resistant.

Proposed Experiments

The evaluation proceeds in four stages, all on the curated lipid data. Stage one reproduces the LANTERN benchmark under our splits and confirms the 0.8161 reference on the transfection endpoint, then measures the same full-feature model on the macrophage endpoint (no reported result found on Lacuna for any predictor's macrophage-uptake accuracy, so we establish it). Stage two runs Lazy-PONSS and all baselines with a fixed training-time budget, reporting per-endpoint $R^2$, subset cardinality, the number of MLP trainings consumed, and LANTERN's own top-candidate ranking precision metric — for which the published number to beat is 73.7% on transfection (LANTERN) and, more ambitiously, a joint ranking that stays high on both endpoints simultaneously. Stage three injects multiplicative label noise via bootstrap resampling of assay replicates at increasing variance, expecting greedy and plain POSS to degrade while PONSS holds its front, mirroring the theoretical collapse proven by Subset Selection under Noise. Stage four is the decisive scientific check: the descriptor-overlap analysis between endpoint-specific fronts, tested against the pre-registered prediction that ionizability descriptors are shared and lipophilicity/packing descriptors diverge — a result that, if confirmed, gives formulation chemists a concrete rule for navigating the evasion-versus-escape tension, and if refuted, remaps where that tension actually lives. Success is declared if a subset at half the full feature count holds within two points of the baseline $R^2$ on both endpoints, the top-lipid joint ranking precision approaches the published 73.7% single-endpoint figure, and the lazy layer cuts trainings by at least an order of magnitude at equal front quality.